Scribd Logo
Upload

Welcome to Scribd!

  • Grassman Algorithm

    Uploaded by

    Fantasy Solution
    17 views2 pages
    The Grassman algorithm partitions face images from video frames into K groups. It samples and characterizes a registration manifold to select frames based on feature richness, which helps extract frames most likely to contain discriminatory features. The algorithm first detects faces, normalizes them, then transforms the video onto a Grassman manifold trajectory linking manifold points. It computes similarities between human joint trajectories to identify people in video sequences.

    Original Description:

    ggggg

    Original Title

    GRASSMAN ALGORITHM

    Copyright

    © © All Rights Reserved

    Available Formats

    DOCX, PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document
    The Grassman algorithm partitions face images from video frames into K groups. It samples and characterizes a registration manifold to select frames based on feature richness, which helps extract frames most likely to contain discriminatory features. The algorithm first detects faces, normalizes them, then transforms the video onto a Grassman manifold trajectory linking manifold points. It computes similarities between human joint trajectories to identify people in video sequences.

    Copyright:

    © All Rights Reserved
    Download as DOCX, PDF, TXT or read online from Scribd
    Flag for inappropriate content
    17 views2 pages

    Grassman Algorithm

    Uploaded by

    Fantasy Solution
    The Grassman algorithm partitions face images from video frames into K groups. It samples and characterizes a registration manifold to select frames based on feature richness, which helps extract frames most likely to contain discriminatory features. The algorithm first detects faces, normalizes them, then transforms the video onto a Grassman manifold trajectory linking manifold points. It computes similarities between human joint trajectories to identify people in video sequences.

    Copyright:

    © All Rights Reserved
    Download as DOCX, PDF, TXT or read online from Scribd
    Flag for inappropriate content
    of 2

    GRASSMAN ALGORITHM:

    For each frame in a video sequence, we first detect and crop the face regions. We then
    partition all the cropped face images into K different partitions. We partition the cropped faces
    by a Grassman algorithm type of algorithm that is inspired by video face matching algorithm.
    Sampling and characterizing a registration manifold is the key step in our proposed approach.
    The proposed algorithm presents a novel perspective towards frame selection by utilizing feature
    richness as the criteria. It is our assertion that quantifying the feature richness of an image helps
    in extracting the frames that have higher possibility of containing discriminatory features. In
    order to compute feature-richness, first the input (detected face) image I is preprocessed to a
    standard size and converted to grayscale. By performing face detection first and considering only
    the facial region, we ensure that other non-face content of the frame does not interfere with the
    proposed algorithm. Given a pair of face coordinates, we determine a set of affine parameters for
    geometric normalization. The affine transformation maps the (x, y) coordinate from a source
    image to the (u,v) coordinate of a normalized image.

    Input: A set of P points on manifold

    {X i }Pi=1 ∈G ( d , D )

    Output: Karcher meanμK

    1. Set an initial estimate of Karcher mean μ K = X iby randomly picking one point in X i }Pi=1

    2. Compute the average tangent vector

    P
    1
    A= ∑ log μK ( X i )
    P i=1

    3. If ‖ A‖<ε then return μK stop, else go to Step 4

    4. Move μK in average tangent direction μK=expμK ( αA ) , whereα >0 is a parameter of step size.
    Go to Step 2, until μK meets the termination conditions (reaching the max iterations, or other
    convergence conditions
    Thus, the video is transformed on a trajectory that links different points on Grassmann
    manifold. The projection on Grassmann manifold requires decomposition. The main advantages
    of this projection are being reversible and have no loss of information. The next step consists on
    similarity computing between human skeletal joint trajectories in order to identify the identity of
    a given skeleton sequence.

    You might also like